The New AI Agents That Shop, Plan, and Gossip

1. Key Themes

Trust Is the Gating Factor, and It Is Quantifiably the #1 Fear

David Pawlan, who has run Assistant Benchmark across 140+ assistants and monitors several AI power-user group chats, quantified what holds the category back. Trust dominates, and it is why finance (the high-value vertical) hasn't materialized. "From over 147 different assistants that have come across my plate so far, the number one fear that people talk about, 22% of the messages are talking about trust as the main fear." [00:00:08] He ties the absence of finance agents directly to this: "The one that I have not really seen come to fruition yet is more of like the finance side of things." [00:05:45] The implication: whoever cracks verifiable trust unlocks the highest-value agent category.

Fences, Not Autonomy: Human-in-the-Loop Is the Right Design for Now

Josh Elman argues the market has the autonomy narrative backwards. Using Shane Mac's "fence in the yard" analogy for the Picnic concept, he says the agent should roam freely inside bounded permissions, and the human should retain the final decision. "I think as humans, we actually don't want to give away that sense of control. I think we actually want it to check in with us, but do all the groundwork, save me all the time in coming up with the best options and the right trade-offs. And now, let me make the final decision." [00:08:44] Anish Acharya adds that rules should be replaced by judgment-based principles: "Instead of giving models rules, we should give them laws and principles." [00:07:59] Josh's cautionary case: a personal agent bled into work and "blasted his finances on his company Slack." [00:07:16]

The Real Trust Risk Is Embarrassment, Not Privacy

Anish separates stated from revealed preferences on privacy: "In revealed preferences consumers will share any amount of data for even the most trivial amount of upside." [00:09:37] David's evidence: Hark offered $30 of free pizza in exchange for access to a user's entire Google Workspace, and people took it. [00:10:04] What consumers actually guard against is social exposure: "The most horrifying visceral thing is that it embarrassed him." [00:10:04] Josh extends it to discretion between friends ("whose agent tells them that you can't eat the cheese?") [00:11:48], and David identifies tone-matching in outbound email (Town) as a core trust layer. [00:12:12]

Agent-to-Agent Social Networks and "Social Indirection" Are the Sleeper Category

Josh describes gossipy agents connected to friends that surface real-life information, such as a friend's power being out, that humans wouldn't volunteer. Anish names the mechanism: "It provides a degree of social indirection that actually makes you closer... something that can create almost a social lubricant in a very sort of low-key way." [00:33:09] The founders of one such product can read the agent-to-agent transcripts but "haven't figured out how to productize that," yet call them "incredibly exciting." [00:32:15] David's meet-up ran a live agent room that devolved into gibberish after about 150 messages without a purpose, while Josh argues goal-directed agents are where discovery happens. [00:38:34]

Judgment Under Conflicting Preferences Differentiates Agents

David's flower test (cheap plus same-day) and LA trip test ("luxurious, cost-efficient weekend") showed that agents resolve tension differently. One thinks cost-efficiency means a red-eye, another treats luxury as time. [00:14:40] In a multi-agent chat room, only Grokbot methodically absorbed other agents' plans into its own. [00:16:49] Anish's insight: using agents forces implicit preferences to be explicit, "you now have to kind of maybe like get over that and tell the agent, 'I'm actually pretty cheap.'" [00:18:22]

Memory, Context, and Portability: Switching Costs Are Lower Than Assumed

Only 3% of 354 surveyed AI power users used Dot as their primary agent, yet the top reason people wanted to try it was built-in ChatGPT context. [00:19:00] But David tested the top 15 agents on the benchmark for memory export, and "all but one were were actually willing to do it." [00:20:17] A common onboarding pattern is now "paste this prompt into ChatGPT, download the file, upload it to me." [00:20:51] Anish's caveat: "We'll see how long that lasts." [00:21:02]

Long-Tail Brands and Direct-to-Consumer Are Structural Winners; Aggregators Get Compressed

Anish: "I actually think the long tail are going to be huge structural winners in this, because they're ones who presumably have better products, but worse marketing, and the agent has the sort of infinite patience to go search for things." [00:24:48] Josh pairs this with Shopify's agent-enablement and the squeeze on aggregators: "Maybe aggregators are the ones that actually get compressed, which is why they're doing so much defensive work here." [00:25:24] Data point from David: 35% of surveyed power users do Amazon/retail shopping through agents, the second most popular category. [00:24:14]

New Data Sources Beyond Email and Messages: Photos, Screenshots, Hardware, Real-World Behavior

The group brainstormed untapped context: camera rolls (Dazzle), screenshots-as-taste (Taste), purchase histories (Amazon, App Store), IoT hardware (a vacuum or AC connected via MCP as passive personality signal), and ambient capture (wearables, Siri Recap). Josh flags the inference problem: "ChatGPT thought I was like a soccer fiend. And I was like, 'No, no, no. I was just into the World Cup.'" [00:45:10] Agents can't yet distinguish curiosity journeys from identity.

Infrastructure and Scale Remain a Real Moat

Contra the "software is a commodity" narrative, David observes: "There really still is some moat there... the agents that are able to perform with hundreds of thousands of users, there are some that are starting to scale up and the performance just degrades immediately." [00:35:14] Browser-use innovation (Kindo's Chrome extension, Hark's visible browser) is a differentiator on task completion, the second-biggest pain point after trust.

The Next Frontier: From Productivity to Aspiration, the "Secret Superpower" Agent

All three converge on the thesis that productivity tooling is the wrong frame. Anish: "I think people want to aspire to be a whole different superhuman version of themselves... I don't know that we have an agent yet that really speaks to those aspirations." [00:48:45] Josh: "When do these agents flip from getting something done to take it off our plate to giving us a secret superpower?" [00:49:18]

2. Contrarian Perspectives

Consumers Won't Be Made More Rational by Agents; Agents Will Have Their Own Irrationality

David cites Patrick Collison's vision of a more rational consumer, and Anish pushes back: "Agents are going to have their own revealed preferences and their own sort of directions of irrationality, though it will not be of a human mind, it will be in of an alien mind." His example: "Agents may have an irrational preference for clean APIs with great docs." [00:26:24] This implies a new discipline of marketing and "consumer research for agents," paid per transaction. [00:26:53] Josh adds that people will reject a paternalistic agent: "I'm actually not as convinced that agents are going to make us more rational." [00:27:30]

Paternalism Kills Products, So the Winning Agent May Serve Our Shadow Side

Anish, drawing on "30 years in consumer product": "Paternalism will kill any product. Even a hint or a whiff of paternalism will... maybe people want the exact opposite of it... I love spending money. I love being perhaps irrationally optimistic about what my future looks like." [00:27:59] Later he calls for "almost like a dark agent" built for the parts of ourselves "that we don't either want to think about, or reveal to perhaps even ourselves." [00:00:17] Most agent startups are starting from the guilt-inducing to-do list.

Stated Privacy Concerns Are Overweighted; Ignore Them in Favor of Revealed Behavior

Contrary to the privacy-first positioning of many products, Anish notes people have linked bank accounts via Plaid "since the Venmo era." David's pizza-for-workspace example supports it. [00:10:11] Josh offers a counterweight: at utility parity, people choose privacy. [00:11:03] The contrarian takeaway is that utility, not privacy architecture, is the primary lever, with privacy as a tiebreaker.

Full Autonomy Is the Wrong Product Goal for Today

The industry narrative is "book me the trip, make me the reservation." Josh says people don't want that: "I like planning a trip. I don't necessarily want you to take care of all of that for me." [00:49:53] The agents that win will do the groundwork and present trade-offs.

Interest-Based Social Matching Probably Needs Hierarchy, Not Just Shared Interest

David is enthusiastic about agents connecting strangers (Tomo's Interest Groups for LSAT studiers). Anish is skeptical: "The interest-based thing has been tried, and there's a little bit of an interesting sort of social hierarchy thing, which is everybody wants to hang out with somebody who feels like they're one step ahead in some direction." [00:41:11] David adds that it works only for people with a desire for new friends, such as transplants. [00:41:59]

3. Companies Identified

Instinct

An AI assistant product on the benchmark, cited as a consumer-facing OpenClaw-style assistant. Mentioned for excellent judgment on the flowers test, "Instinct really did continuously bring back the best price." [00:13:49] It sat "right down the middle" on cost vs. time in the LA trip test [00:15:36], and David's technophobic mom adopted it easily: "She's having so much fun... It's like a kid in a candy shop." [00:47:20] Anish also cites it (with Muse) as a consumer version of OpenClaw. [00:37:41]

Kindo (referred to as "Kicker" later in the transcript)

An agent that has users install a Chrome browser extension. Mentioned for browser-use advantage on the flowers test: "It has you install a Chrome browser, and so its browser use has a little bit of an advantage from that point." [00:13:49]

Ceph

A benchmarked assistant. "Ceph also did a really good job" on the flower test, balancing price and delivery time. [00:13:46]

Grokbot

An assistant that stood out in the multi-agent chat-room test: "Grokbot was actually the only one that, you know, very methodically took value from the other ones and implemented into its own plan." [00:16:50] It leaned toward prioritizing time in the LA trip test. [00:15:36]

Dot

An assistant with built-in ChatGPT context. It leaned toward prioritizing cost in the trip test [00:15:36], and was used as the primary agent by only 3% of 354 power users, though the context advantage is its pull. [00:19:00]

Town

Praised for tone-matching in outbound email: "Something that Town has done remarkably well is really nailing the tone, and building that trust layer of being able to send out emails on your behalf." [00:12:11]

Wabi

A multi-thread agent product that just launched. Anish says it creates "a new kind of blank piece of paper where you can establish your social contracts together" when friends and agent are all onboarded at day zero. [00:31:07] Josh uses it for thread-based life management (watch lists, books) with widget UIs. [00:30:12] David sees it as the multiplayer play that "becomes a little bit more like entertainment." [00:29:13]

Scout

A shopping agent adding "a unique kind of like swipe-type flavor to learn your preferences and adds a little bit more entertainment there." [00:29:13]

Hark

An agent with a large marketing campaign offering $30 of free pizza in exchange for Google Workspace access. It also has a visible, fast-looking browser UI. [00:10:53] [00:34:35]

Tomo

Has an "Interest Groups" feature that clusters users (e.g., LSAT studiers) into group chats where agents can connect them. [00:39:47]

Dazzle (Marissa Mayer's new company)

Photo-roll-based personal context product with a forgetting-by-default privacy thesis: "If you don't need it, they're going to forget about it." [00:22:15] [00:22:15] Not yet added to the benchmark but near the top of David's queue.

Picnic (Shane Mac's concept)

A trust and permissioning framework using fences. [00:06:55]

Taste (working name)

An early-stage founder's product that organizes saved screenshots into mood boards and feeds them to code generation as a style profile. [00:23:42]

Shopify

Cited as empowered by agentic commerce because it backs direct brands and is making its catalog available to agents. [00:25:20]

OpenClaw and Moltbook

OpenClaw was the home-Mac-Mini-based assistant paradigm that "started all of this." Moltbook was the agent social network. Anish: "What is the mass market version of Moltbook? ... Something like that will exist. How can it not?" [00:37:41]

Thinking Machines

Referenced for letting users tinker with their own models. "If we can eventually hit a world where that takes into the consumer form in a really easy way, it's going to be an interesting space." [00:36:59]

Wejo (as transcribed; hardware-MCP provider)

Just released an MCP enabling connection to hardware such as AC units and vacuums, which David believes can produce passive personality context. [00:42:52]

Apple

Cited for its privacy positioning and private cloud compute, and Apple Watch features (Siri Recap, live rewind) that Josh upgraded for. [00:11:03] [00:46:21]

Amazon Bee

Described as a "24/7 ambient listener just wearing on your wrist" in the ambient capture discussion. [00:46:31]

Plaid

Used as evidence that consumers connect financial accounts readily for modest utility. [00:09:46]

Amazon and Delta

Named as examples of companies restricting agent access. [00:24:44]

Underdog (as transcribed)

Mentioned as an on-device assistant exploring local models and infrastructure. [00:36:34]

4. People Identified

Shane Mac

Founder behind the Picnic concept. His fence analogy and his experience of a personal agent leaking his finances into company Slack made him the case study in agent trust. Josh: "He's been talking about this concept of something called Picnic." [00:06:49]

Patrick Collison

Stripe co-founder. David cites his tweet about a future of more rational consumers as agents do the work, and Anish praises the vision while pushing back. "I love Patrick's vision." [00:26:09]

Marissa Mayer

Founder of Dazzle. Praised for emphasizing the trust layer in a photo-based agent. [00:21:36]

David Pawlan

Creator of Assistant Benchmark, who has manually tested 140+ assistants, hosted a 100-person agent meetup, and runs surveys of power users (354 respondents). He provides the empirical backbone of the episode, such as "22% of the messages are talking about trust." [00:00:08]

Anish Acharya

a16z partner with 30 years in consumer product. He built a screenshot-extraction app, signs AI-assisted emails with the model name as a transparency signal [00:12:46], and offers the paternalism and "dark agent" frameworks. [00:27:54]

Josh Elman

Investor. A heavy user of gossipy agents, Wabi, and Tomo, and the source of the "fence" and "secret superpower" framings. [00:49:17]

David's mom

An anecdotal bellwether for mainstream adoption: a self-described technophobe who struggled with ChatGPT but enjoys texting an AI assistant, and used it to find and negotiate with a breeder for a dog. [00:47:41]

5. Operating Insights

Disclose AI Use With a Signature to Preempt "Hoodwinked" Reactions

Anish's tactic: "I sign all the email that I write with AI with the model that I used. I write my own name plus the model number so that I'm just signaling to the person like, 'Hey, I'm acknowledging that I used AI for this.' Cuz I think actually people are pretty flexible as long as they don't feel hoodwinked by it." [00:12:38] For founders shipping outbound-email agents, this suggests a default disclosure pattern that reduces the embarrassment risk.

Design Onboarding Around Context Import and Conversational Interviews

Two onboarding patterns emerged. One has the product give users a prompt to run in ChatGPT, then upload the memory file. [00:20:51] The other is a voice call where the agent asks questions. Josh: "They're very disarming. They just get you talking." [00:21:18] Both bypass the cold-start problem and exploit the fact that most agents (14 of 15 tested) will export memory.

Make Trade-Off Elicitation a Core Feature, Not an Afterthought

Josh's critique of the trip test: the best agent behavior is to surface tensions and present options (red-eye plus hotel nap vs. cost). "Getting to know the individual person and break down those tensions and present the trade-offs in different options is really important." [00:16:06] Operationally, build clarifying-question flows and "second-layer dimension" prompts, such as asking whether to add a gift note or whether a recipient wants more than cheapest.

Use Group Threads With a Clean Slate to Establish Social Contracts

Anish: when friends and an agent start a thread simultaneously, "everybody's on an equal playing field on D0," establishing norms from scratch, whereas adding an agent to an existing iMessage thread feels invasive. [00:31:07] Josh's caution is to avoid duplicate threads with the same people and keep each thread topic-scoped so trust boundaries are clear. [00:31:44]

Test Scale Before You Scale Marketing

David's observation that performance degrades immediately for products once they hit hundreds of thousands of users, with founders dealing with outages, implies that load testing and infrastructure reliability are a competitive moat in this phase. [00:35:14]

6. Overlooked Insights

Agents' Cross-Pollination of Plans Is a Product Insight Hiding in a Benchmark Detail

David noted offhand that when he put three agents in a chat room over three rounds, "Grokbot was actually the only one that, you know, very methodically took value from the other ones and implemented into its own plan." [00:16:50] This is a quiet but important capability: the ability to learn from peer agents in-context. In a world heading toward agent-to-agent networks, the agent that can absorb and integrate peer information, without being derailed (as others "spoke gibberish" after 150 messages in the freeform room), may be the one that wins in any marketplace or discovery network.

Founders Can Read Agent-to-Agent Transcripts but Haven't Productized Them

Josh mentions that the founders of a gossipy agent "can read some of the transcripts between the agents... they haven't figured out how to productize that." [00:32:15] This is a nearly unexplored data asset: a record of how agents negotiate on behalf of humans, effectively a proxy for latent social preference and relationship graphs. Whoever figures out how to turn that into a product (matchmaking, introductions, serendipity) has a novel social primitive.

Agents May Need Their Own "Consumer Research" Layer Paid Per Transaction

In passing, Anish suggests that "we'll need some sort of a consumer research, but for agents, that will then presumably be paid per transaction to do the research." [00:26:53] This implies a new intermediary category (agent-facing Wirecutter or review aggregator) with a business model tied to transactions instead of ads, because agents, unlike humans, can't rely on brand or marketing as a quality shortcut.